model: Add support for Tiny Aya Models (#19611)

* changes for tiny aya

* changes to hash

* changes to vocab

* fix some tokenizer regex edge cases

* update comment

* add some comments for regex

* Apply suggestion from @ngxson

---------

Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com>

src/llama-vocab.cpp

@@ -422,6 +422,14 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_TINY_AYA:
+                regex_exprs = {
+                    // original regex from tokenizer.json: "\\d{1,3}(?=(?:\\d{3})*\\b)"
+                    "\\d{1,3}(?=(?:\\d{3})*\\b)",
+                    // original regex from tokenizer.json: "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+"
+                    "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             case LLAMA_VOCAB_PRE_TYPE_KIMI_K2:
                 regex_exprs = {
                     // K2 trigger pattern - this will activate the custom K2 handler in unicode.cpp
@@ -2005,10 +2013,14 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "megrez") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
             } else if (
-                    tokenizer_pre == "gpt-4o" ||
-                    tokenizer_pre == "llama4") {
+                tokenizer_pre == "gpt-4o" ||
+                tokenizer_pre == "llama4") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "tiny_aya") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_TINY_AYA;
+                clean_spaces = false;
             } else if (
                 tokenizer_pre == "superbpe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_SUPERBPE;

src/llama-vocab.h

@@ -55,6 +55,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_YOUTU           = 44,
     LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE      = 45,
     LLAMA_VOCAB_PRE_TYPE_QWEN35          = 46,
+    LLAMA_VOCAB_PRE_TYPE_TINY_AYA        = 47,
 };
 
 struct LLM_KV;

src/unicode.cpp

@@ -769,6 +769,12 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
     } else if (regex_expr == "\\p{AFMoE_digits}") {
         // AFMOE digit pattern - use custom implementation for proper splitting
         bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
+    } else if (regex_expr == "\\d{1,3}(?=(?:\\d{3})*\\b)") {
+        // tiny_aya digit grouping pattern from tokenizer.json:
+        //   {"type": "Split", "pattern": {"Regex": "\\d{1,3}(?=(?:\\d{3})*\\b)"}, "behavior": "Isolated"}
+        // Splits digits into groups of 3 from the right (e.g., 1234567 -> 1, 234, 567)
+        // TODO: Revisit this regex, incase there are any subtle tokenization differences with the original regex.
+        bpe_offsets = unicode_regex_split_custom_afmoe(text, offsets);
     }
 
     return bpe_offsets;